Post on 05-Jul-2020
A Company’s Digital TwinTUM Data Innovation Lab
Team: Sagarika Kathuria, Pooreumoe Kim,Frederik Wenkel, Jieyi Zhang
Mentors: Simon Brand, Anton Kurz, Sebastian Rossner
Co-Mentor: Laure Vuaille
Project Leader: Dr. Ricardo Acevedo Cabra
Supervisor: Prof. Dr. Massimo Fornasier
1
Our Team
2/37
Sagarika KathuriaMathematics in Data Science
Pooreumoe KimData Engineering and Analytics
Frederik WenkelMathematics
Jieyi ZhangData Engineering and Analytics
AGENDA
02
03
04
Motivation
Digital Twin Concept
Process Mining at Celonis
Methodology
05
06
Specific Use Case at Celonis
Summary & Outlook
01
3/37
AGENDA01
02
03
04
Motivation
Digital Twin Concept
Process Mining at Celonis
Methodology
05
06
Specific Use Case at Celonis
Summary & Outlook
4/37
Company as a Complex System
5/37
Resources Market
Company
➔ Company’s future performance?➔ Forecasting systems
➔ Sales forecasting, predictions for volatile resources
➔ “Patchwork” of forecasts➔ No complete picture
Competitors
AGENDA
02
03
04
Motivation
Digital Twin Concept
Process Mining at Celonis
Methodology
05
06
Specific Use Case at Celonis
Summary & Outlook
6/37
01
Digital Twin Concept
7/35
Digital Twin Concept
8
Major characteristics
➔ Similar behavior to original organization
➔ Family resemblance
➔ No exact copy➔ Digital Twin can be
raised differently
AGENDA
03
02
Motivation
Digital Twin Concept
Process Mining at Celonis
Methodology
05
06
Specific Use Case at Celonis
Summary & Outlook
01
9/37
04
Process Mining at Celonis
10/37
➔ Every process leaves a digital footprint in the company’s environment
➔ Celonis Process Mining enables an organization to structure its raw data of a process into a data model
➔ The main object : Activity Table
➔ Consists of a number of activities represented by cases
➔ Each case has its own case key to distinguish different cases
➔ Each Activity has an event time,a sorting and a category
ACTIVITY TABLE
Process Mining at Celonis
11/37
➔ Analysis in Celonis Data Mining:
➔ Helps visualize and understand the data using tools like Process Query Language(PQL) and Key Performance Indicators (KPI’s)
➔ Helps identify bottlenecks and inefficiencies within the process
➔ Consists of various objects like: Process Explorers, Variants Explorer, Standard KPI's
➔ Customizable with tables, charts, diagrams and custom KPI’s ANALYSIS
AGENDA
04
02
03
Motivation
Digital Twin Concept
Process Mining at Celonis
Methodology
05
06
Specific Use Case at Celonis
Summary & Outlook
01
12/37
Explorative Data Analysis
13/37
➔ Input Dataset : Celonis Happyfox IT Service Management
➔ Happyfox - a Saas Platform, offers help deskticketing system for approx. 12000 companies.
➔ Celonis Happyfox ITSM Data Model:
➔ Three data tables which contains the information for each ticket - _CEL_ITSM_happyfox_ACTIVITIES, Tickets, Updates
➔ Data stored in the form of string, boolean, number and date time values
➔ Data accessed using PQL (Process Query Language) in Celonis and Python API.
Explorative Data Analysis
14/37
➔ ITSM Ticket Processing
➔ Each ticket - has its own process - each process type called a Variant
➔ Happy Path - variant that happens most frequently
➔ Throughput time -
Ttp
= tsact
− tsact_previous
Ttotal_tp
= tslast_act
− tsfirst_act
➔ Total Variants : 1850 , 20% cases follow Happy Path
Explorative Data Analysis
➔ Throughput Time Analysis
➔ 2 cases - tickets created in 2017 & 2018
➔ Median - close to 0 daysMax Throughput Time - 276 days
➔ Digital Twin model considers outliers as well
2018
Throughput Time (days) 15/37
17/18
2017
Throughput Time (days)
Explorative Data Analysis
Throughput Time Analysis
➔ Kolmogorov–Smirnov(KS) Test
➔ Tests the distance between the empirical distribution function of the data and the cumulative distribution function (CDF) of the reference distribution.
➔ H0 = Two distributions follow the same
distribution
➔ H0 holds if p-value > 0.05
➔ Need non-parametric methods to simulate the Data such as Bins Method, Kernel Density Estimation
16/37
Distribution p-value Is passed
Poisson Distribution
4.675e-10 No
Exponential Distribution
0.0 No
Birnbaum-Saunders
0.0 No
KS Test Results
Explorative Data Analysis
17/37
➔ Category Based Analysis
➔ Total Categories - 21
➔ Categories of Importance - 5
➔ Digital Twin simulates categories in cases based on their percentages
Count of Tickets based on Categories
Explorative Data Analysis
18/37
➔ Ticket Number Analysis
➔ Total Tickets - 12000
➔ Dickey Fuller Test - Statistical test for checking stationarity.
H0 - Series is non-stationary
p-value : 0.091Cannot reject H
0as p-value > 0.05
➔ Use Transformations like log
➔ Estimate trend and seasonality to predict number of tickets using Time Series Methods
Count of Tickets Based on Creation Date
19/37
General Simulation Approach
Act
iviti
es
Even
t Ti
mes
Cat
egor
ies
Input
Activity Simulation
Category Simulation
Event Time Simulation
Output
1
3
2
➔ First goal: Family resemblance
➔ Second goal: Raise twin differently
Activity Simulation
20/37
➔ First building block of simulation➔ Importance for subsequent blocks
➔ Measure of precision: Occurrences of most frequent activity flows from input data in output
➔ According to empirical observations from input
➔ Markov Process
➔ Manageable matrix representation➔ No dependencies captured on activity
history
21/37
Activity Simulation
➔ Preprocessing yields improvement➔ Treat activity flows with different starting activities separately
➔ Linear Additive Markov Process (LAMP) instead of ordinary Markov Process
➔ Parameters w, P have to be learned minimizing the following negative log likelihood
Category Simulation
22/37
➔ Category assignment to every activity flow from activity simulation
➔ Starting category assignment➔ Markov Process for category changes Category 1
Category 2
The number of Tickets Prediction
23/37
❏ Train time series models from training set
❏ Data Separation
❏ ARIMA, Holt Winter’s ...
❏ Compare the prediction result to test set
❏ Measure errors
The number of Tickets Prediction
24/37
➔ Choose Holt Winter’s Model
Throughput Time Simulation: KDE
25/37
❏ Kernel Density Estimation(KDE):Non-parametric way to estimate the probability density function(PDF) of a random variable.
❏
❏ Two important parameters
h : Bandwidth K(.) : Kernel function
Throughput Time Simulation: KS-Test
26/37
❏ Kolmogorow-Smirnow-Test (KS-Test)
❏
❏ p-value > 0.05:
Two sets have same distribution
The higher, the more identical
Throughput timeFrom <Status: New> to <Change assignee>
H0: FX(x) = FY(x)H1: FX(x) ≠ FY(x)
Throughput Time Simulation: Kernel Selection
23/37➔ Choose Uniform Kernel
❏ Simulating power(KS-Test)?
❏ Identical
❏ Table 7 (Doc)
❏ Negative value generation?
❏ Except Gaussian, no negative if
bandwidth <= 0.1
❏ Table 8 (Doc)
❏ Internal sampling method?
❏ Gaussian & Uniform
Throughput Time Simulation: Bandwidth Selection
28/37
❏ Three methods for bandwidth:❏ Constant bandwidth (0.1)
❏ Gridsearch
❏ Gridsearch with Cross validation
❏ Analysis of variance(ANOVA)
Comparison of running times
ANOVA result: cannot reject H0
➔ Choose Constant Bandwidth(0.1)
H0 : µ1 = µ2 = µ3
H1 : µi ≠ µj
Digital Twin: Integration of the Methodologies
29/37
❏ Activity Simulation❏ Category Simulation❏ Number of Cases❏ Throughput Time
➔ Generates Virtual Activity Table
Model Validation
30/37
ModelMarkov Process
KDE (Bandwidth=0.1)
HoltWinter(Period = 7 days)
M1 2 months 2 months All
M2 2 months 2 months 4 months
M3 1 month 1 month 4 months
● Training Data
● Cross Validation
Model Validation
31/37
● Simulated Values Validation (Prediction for May 2018)
Cases per day Events per Day Avg Total Throughput time
Trimmed Avg Total Throughput Time
Sample Size
rel. Error M1 31.25% 28.57% 15.52% 16.05% 21.45%
rel. Error M2 25.00% 32.38% 11.20% 11.11% 23.55%
rel. Error M3 18.75% 20.95% 18.10% 14.81% 25.45%
● Simulated Values Validation (Prediction for June 2018)
Cases per day Events per Day Avg Total Throughput time
Trimmed Avg Total Throughput Time
Sample Size
rel. Error M1 45.16% 36.99% 43.53% 33.33% 13.40%
rel. Error M2 38.71% 30.64% 42.35% 31.82% 14.87%
rel. Error M3 36.67% 27.75% 28.24% 18.18% 14.87%
Model Validation
32/37
● Model 3 Total Throughput Time Distribution (2018-05) Activity Frequency Validation (2018-05)
Create Ticket 21.34% 21.36%
Status: New 16.46% 17.95%
Change assignee 12.20% 9.40%
Status: Open 11.59% 12.82%
Status: Closed 10.37% 9.40%
Change Category 9,76% 9.40%
Status: Solved 7.93% 6.84%
Status: On Hold 5.49% 6.84%
Change Priority 4.89% 6.84%
0.09
9.05
22.95
10.61
9.35
3.69
13.75
24.59
39.87
Activity FrequencyReal Data
FrequencyPrediction
Rel. ErrorIn %
+
+
Real Data Total Throughput Time (in Days)
Prediction Total Throughput Time (in Days)
AGENDA
02
03
Motivation
Digital Twin Concept
Process Mining at Celonis
Methodology
06
Specific Use Case at Celonis
Summary & Outlook
01
33/37
05
04
First Level Service Automation
34/37
Reduce the throughput time of steps which belong to first level service to zero and simulate the cases.
Simulated Value for May 2018
Cases per day Events per Day Avg Total Throughput time
Trimmed Avg Total Throughput
Time
Sample Size
Percentage of Change -6.25% -4.76% -68.10% -78.75% -24.09%
What if I buy a chatbot and automate the first-level service?
How does this affect my throughput time?How much people could I reallocate?...
First Level Service Automation
35/37
AutomatedReal World
● More tickets solved within a short time
● Average throughput time overall reduced
● Same Happy Paths
Tickets Created in May 2018 (Celonis Process Explorer)
AGENDA
02
03
Motivation
Digital Twin Concept
Process Mining at Celonis
Methodology
Specific Use Case at Celonis
Summary & Outlook
01
36/37
06
04
05
Summary and Outlook
37/37
❖ Process mining and Digital Twin Model
❖ Happyfox ITSM Data Model
❖ Data exploration
❖ Introduction of the techniques applied
❖ Model training and validation
❖ Application of Digital Twin
❖ More sophisticated model
❖ More what-if questions
➢ Number of tickets change
➢ 24/7 customer support
❖ More input data
❖ ...
Summary Future Works
Thank you!